Question Qualcomm's first Nuvia based SoC - Hamoa

[poke01]

Qualcomm's working on a 2024 PC chip codename "Hamoa" with up to 12 (8P+4E).

Said to the same cache layout as M1. large private L1$, per-cluster L2$ (cluster = 4 cores, 12MB for every cluster) and a lot of LLC.

source:

https://twitter.com/x/status/1589405172979339264

 

[FlameTail]

Rumour: Qualcomm, Samsung working with AMD on FSR

https://www.reddit.com/r/Amd/comments/17mrf1u

Interesting developments. The way I am interpreting this rumour is that Samsung wants to have FSR on their Galaxy smartphones. Which is obvious since Exynos processors* now use AMD's RDNA GPU IP. The reason Qualcomm is involved is since Samsung dual-sources both Exynos and Snapdragon chips for their flagship S series. They want one single solution working on their phones: FSR.

Still this does leave open for further possibilities, as the commenters have noted. Will Qualcomm and AMD further collaborate on FSR, with the potentiality of FSR coming to future Snapdragon X processors ? On one hand FSR is inferior to DLSS, so it could benefit from more engineers working on it to improve it. On the other hand, Snapdragon X is a direct threat to AMD's Ryzen, so it may be detrimental for AMD to collaborate.

 

[moinmoin]

FSR is open source, so anybody can use it even without collaborating with AMD. Plenty games already use it, even on Nintendo Switch which contains zero AMD hardware.

 

[FlameTail]

OCL -> Rusticl -> Zink -> Vulkan.

Hmm. QC already has OpenCL drivers on Android. Can they not simply be ported over to Windows?

 

[soresu]

Hmm. QC already has OpenCL drivers on Android. Can they not simply be ported over to Windows?

Dunno.

IMHO at the moment no one has really good OCL drivers for Windows so it likely won't matter much anyway.

AMD made a poor effort on their driver, and nVidia pretty much kneecapped theirs from day 1 to prevent it looking like a viable CUDA alternative and thus damage their carefully engineered market monopoly.

 

[DrMrLordX]

OpenCL is mostly deprecated anyway, isn't it?

 

[soresu]

OpenCL is mostly deprecated anyway, isn't it?

Certainly OCL 3 didn't sound like a step in the right direction to me.

It felt more like a Khronos capitulation and endorsement of nVidia's "I won't support anything past v1.2" strategy to artificially keep OCL marketshare low vs CUDA.

 

[FlameTail]

So if OCL is dead, then what are the replacement/alternatives?

 

[FlameTail]

This Adreno GPU in the X Elite is pulling a lot of power. Over 30 watts.

Compare that to their mobile SoCs:



8gen2 pulls 7.5 watts. 8gen3 pulls 10.7 watts.



X Elite: 44 FPS @ 33W
SD8G2: 22 FPS @7.5W
SD8G3: 32 FPS @10.7W

According to some rumours from Chinese forums, the X Elite's iGPU is based on the architecture of the SD8G2, and has approximately 2x the ALUs. Well, that tracks with the 2x FPS number. But what isn't tracking is the power consumption which increased by 4x.

Even if we doubt the validity of this rumour and assume the X Elite has the same iGPU as the SD8G2 but now clocked 2x higher, the graph Qualcomm has showed doesn't suit that assumption. Because then the graph would look a lot more flatter as it is pushed out of it's efficiency range.



*Found FPS numbers of 8g2 and 8g3 3DMark WLE from a Google image search.

 

[soresu]

View attachment 88989

Da hell is the ARM Mali solution beating A17 Pro?! 😅

We're in cloud cuckoo land now! 😂🤣😆

 

[igor_kavinski]

Da hell is the ARM Mali solution beating A17 Pro?! 😅

The A17 Pro is at a serious bandwidth disadvantage in that chart.

 

[FlameTail]

The A17 Pro is at a serious bandwidth disadvantage in that chart.

Bahndeith diavantaag ?

 

[FlameTail]

Ah yes.

LPDDR5-6400 vs LPDDRT-9600

51 GB/s vs 76 GB/s

 

[DrMrLordX]

Certainly OCL 3 didn't sound like a step in the right direction to me.

It felt more like a Khronos capitulation and endorsement of nVidia's "I won't support anything past v1.2" strategy to artificially keep OCL marketshare low vs CUDA.

So if OCL is dead, then what are the replacement/alternatives?

Since this thread doesn't really pertain to OpenCL or heterogeneous processing, I'll redirect to the last known related thread I follow on these forums:

Speculation: SYCL will replace CUDA

I've had my eye on SYCL for some time now. I don't think it is widely recognised that this is going to be the standard for heterogeneous programming models going forward (eventually to be merged into the ISO C++ standard, it seems). If you read any discussion on programming, tech and investor...

forums.anandtech.com

 

[qmech]

X Elite: 44 FPS @ 33W
SD8G2: 22 FPS @7.5W
SD8G3: 32 FPS @10.7W

I'm assuming you aren't simply taking the 33W from one graph and the 44 fps from a different table and combining the two. Without giving a link to whatever source you have the "44 fps @ 33 W" claim from, that's what it looks like you are doing.

 

[FlameTail]

I'm assuming you aren't simply taking the 33W from one graph and the 44 fps from a different table and combining the two. Without giving a link to whatever source you have the "44 fps @ 33 W" claim from, that's what it looks like you are doing.

44 FPS from Geekerwan, 33W from Qualcomm!

Dual sourcing! You know it's never a good idea to rely on solely one source, right? /s

 

[FlameTail]

In all seriousness though, both numbers originate from Qualcomm.


33W from this graph.


44 FPS from the Performance Preview Qualcomm held for the X Elite.

 

[qmech]

In all seriousness though, both numbers originate from Qualcomm.

View attachment 89004
33W from this graph.

View attachment 89006
44 FPS from the Performance Preview Qualcomm held for the X Elite.

So you took an arbitrary endpoint from the power curve and decided that's the exact wattage used to achieve the performance listed in a different table.

The 33 W endpoint was chosen for the sole reason that it maximized the delta between the X Elite and the 7940HS. Had Qualcomm marketing chosen to continue the power-performance curve to 50 W, would you have used that number instead?

I may have missed it, but there seems to be nothing that ties the 44 fps to the 33 W.

 

[SpudLobby]

Doug how do you feel about Microsft and Windows on Arm now?

Microsoft is finally making custom chips — and they’re all about AI

Microsoft joins the AI hardware race with two custom silicon chips

www.theverge.com

And before you say: oh that’s just servers.

And? Still counts insofar as Windows server is a thing — see Ampere instances. I mean that’s pretty obviously a clue as to their intentions. Microsoft is not down and out on Arm.

 

[FlameTail]

According to rumours and speculation floating online, the Hamoa chip was supposed to be a 2022 chip, with devices arriving in 2023.

However as things have turned out, the X Elite has been announced at the tail end of 2023, with devices launching in 2024.

2024 seems like a turbulent year. Intel Meteor Lake will be out, and AMD are launching Zen5 and all those formidable APUs.

 

[FlameTail]

Also in other news, Qualcomm has announced the Snapdragon 7 gen 3.

Hah- is this the only major thread regarding Qualcomm? We don't have a thread for discussing about Snapdragon mobile SoCs. Shall I make one?

 

[poke01]

According to rumours and speculation floating online, the Hamoa chip was supposed to be a 2022 chip, with devices arriving in 2023.

However as things have turned out, the X Elite has been announced at the tail end of 2023, with devices launching in 2024.

2024 seems like a turbulent year. Intel Meteor Lake will be out, and AMD are launching Zen5 and all those formidable APUs.

I feel like Lunar lake will shake up the industry like M1 did. Ultrabooks and handheld PC gaming will benefit

 

[FlameTail]

Like- there is literally an Apple Silicon SoC thread. I am jealous now.

Edit: Qualcomm Snapdragon Thread has been created!

 

[FlameTail]

I feel until Lunar Lake will shake up the industry like M1 did.

Not sure what you mean. Is that a grammatical error?

 

[poke01]

Not sure what you mean. Is that a grammatical error?

Fixed

 

[eek2121]

Hmm. QC already has OpenCL drivers on Android. Can they not simply be ported over to Windows?

I don’t have experience with modern (Win 7/8/10/11) windows driver development, but Windows drivers are designed in a fundamentally different way compared to Linux/Android. You cannot simply “port” a driver between Linux/Windows.